Video outlier or anomaly detection is currently performed using computer vision and image processing techniques such as filtering, motion estimation, etc. that require decompressed video streams. Applying machine learning techniques directly to a video stream is computationally expensive. This disclosure describes the use of a pre-trained machine learning model that can analyze compressed video bitstreams directly without decompression to identify likely anomalies in the video. Such direct analysis of compressed video streams can be performed at significantly lower computational cost while still yielding a strong anomaly detection signal. Sequences in which anomalies are detected can be decompressed and further analyzed using conventional techniques of video analysis. Sequences over longer time spans (and associated groundtruth information about whether an anomaly exists in the sequence) can be used for continuous unsupervised training of the same model. This can improve detection of anomalies in subsequent sequences to enable better anomaly detection, without requiring collection of additional training data or separate training cycles.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.